xAI has entered the frontier coding race with a distinctly economic argument. Released on July 8, Grok 4.5 is the company's new flagship for "code and everything else" — a 1.5-trillion-parameter Mixture-of-Experts model that Elon Musk pitched as Opus-class capability delivered faster, cheaper and with dramatically fewer tokens. Rather than claiming outright benchmark supremacy, xAI is betting that efficiency is the metric that will actually decide which model developers keep open in their editors.
Built on V9, Trained Alongside Cursor
Grok 4.5 sits on xAI's 1.5-trillion-parameter V9 foundation and ships with a 500,000-token context window, throughput of roughly 80 tokens per second, and a per-call reasoning-effort dial that lets developers trade depth for speed on each request. Pricing lands at an aggressive $2 per million input tokens and $6 per million output tokens — a fraction of what rival frontier coding models charge.
The model's most unusual ingredient is its training partnership with Cursor, the AI coding editor that SpaceX agreed to acquire in a reported $60 billion deal in June. Grok 4.5 was trained alongside Cursor and on interaction data from real coding sessions, giving xAI something competitors lack: a direct pipeline from a hugely popular development environment into model training. The model is available in Grok Build, across all Cursor plans and via xAI's console, though EU availability was still pending at launch, targeted for mid-July.
The Benchmark Picture: Strong, Selective, Contested
xAI's published results position Grok 4.5 near the top of the terminal-agent leaderboard. On Terminal-Bench 2.1, it scored 83.3 percent — second only to Claude Fable 5's 84.3 percent and ahead of Claude Opus 4.8 at 78.9 percent. On DeepSWE 1.0 it posted 62.0 percent, again clearing Opus 4.8, and it leads SWE Marathon outright at 29.0 percent pass@1.
The fuller picture is more mixed:
- On SWE-Bench Pro, Grok 4.5 scores 64.7 percent, well behind Claude Fable 5's 80.4 percent and slightly below Opus 4.8.
- On DeepSWE 1.1 under the mini-swe-agent harness, it trails all major rivals at 53 percent.
- Independent aggregator Artificial Analysis places it fourth overall on its Intelligence Index at 54 — while ranking it first on agentic tool use.
Analysts noted that Grok 4.5 beats Opus 4.8 on two of the four headline benchmarks xAI published and loses the other two. Two disclosure caveats also drew attention: Cursor acknowledged that a snapshot of its own codebase was accidentally included in training, potentially inflating CursorBench results, and xAI labeled competitors' effort modes (max, xhigh) without stating which configuration produced Grok's own scores.
Efficiency as the Real Headline
The number xAI most wants developers to see is not a score but a token count. On SWE-Bench Pro tasks, Grok 4.5 resolves problems using an average of 15,954 output tokens, versus 67,020 for Claude Opus 4.8 at max effort — a 4.2x efficiency gap. Combined with $6-per-million output pricing, that compounds into an order-of-magnitude cost difference per resolved task for many workloads.
For agentic coding, where models can loop through dozens of tool calls per task, token efficiency translates directly into latency and bill size. A model that reaches a similar answer with a quarter of the tokens is not merely cheaper; it is faster to iterate with, easier to run in parallel and less likely to blow through context windows on long-horizon tasks.
Why It Matters
Grok 4.5 marks the moment the frontier coding race split into two competitions: one for peak capability and one for capability per dollar. Anthropic's Fable 5 still owns the hardest software-engineering benchmarks, but xAI is attacking the much larger territory of everyday coding work, where "good enough, four times cheaper" is a compelling pitch — especially with distribution built directly into Cursor's massive user base. The vertical integration angle deserves equal attention: a frontier lab training on telemetry from a coding editor it is acquiring is a preview of how model quality and product data may consolidate into closed loops that startups cannot easily replicate.
The launch also sharpens an industry-wide credibility problem. Selective benchmark publication, undisclosed effort settings and accidental training contamination are becoming recurring themes across frontier releases, pushing independent evaluators into an increasingly central role. For developers, the practical takeaway is simpler: with Grok 4.5 at $2/$6, GPT-5.6's tiered pricing and open-weight challengers closing in, the cost of frontier-grade coding assistance is collapsing — and the labs know the next battle will be won on efficiency, not just leaderboards.
